[IRMover] Don't map globals if their types are the same

[pirama]

Summary

During IR Linking, if the types of two globals in destination and source
modules are the same, it can only be because the global in the
destination module is originally from the source module and got added to
the destination module from a shared metadata.

We shouldn't map this type to itself in case the type's components get
remapped to a new type from the destination (for instance, during the
loop over SrcM->getIdentifiedStructTypes() further below in
IRLinker::computeTypeMapping()).

Fixes PR40312.

[tejohnson]

llvm/lib/Linker/IRMover.cpp

765

Can you confirm how the DGV was added to the DstM originally? Do you mean it was already imported from SGV based on information in the summary? We don't directly add from the module summary, it is just used to compute the list of GVs to import from a source here via the IRMover.

[pirama]

Fix typo reffers -> refers

llvm/lib/Linker/IRMover.cpp

765

Yes, even without this change, DGV was added to DstM.

Here's the path that I traced inside a debugger:

Inputs/type-mapping-bug3.ll: function 'a' has attached debug metadata !6
Parsing that metadata ends up at metadata !5 (which is shared between both input files)
In type-mapping-bug3.ll, processing !5 leads to !7 which is a ConstantAsMetadata, with value as the function declaration.

This value/declaration gets added to DstM in Mapper::mapSimpleMetadata in ValueMapper.cpp.

[pirama]

llvm/lib/Linker/IRMover.cpp

765

Btw, this (GlobalValue from SrcM already added to DstM) happens for PR37684 (test/LTO/X86/type-mapping-bug2.ll) as well. It's just that the types get appropriately mapped if this type and component types don't involve opaque types. This fix is to handle the case where opaque types are involved.

Also, skipping the mapping here is at worst a delayed optimization since they should get resolved appropriately via TypeMapper::get.

[tejohnson]

llvm/lib/Linker/IRMover.cpp

765

I still don't understand the comment about the module summary, which is just telling the IRMover which symbols to move. What are DGV and SGV in the case where the types match?

llvm/test/LTO/X86/Inputs/type-mapping-bug3.ll

9

Second sentence is incomplete.

llvm/test/LTO/X86/type-mapping-bug3.ll

41

Comment seems stale - there is no !11

45

%t0.o?

[pirama]

Update/clean-up comments in the test.

[pirama]

llvm/lib/Linker/IRMover.cpp

765

I incorrectly used "module summary" in the patch. The value is getting loaded via a shared metadata node (!5 in both test inputs)

[pirama]

Ping...

@tejohnson Did my reply about shared metadata clarify your question?

[tejohnson]

In D66814#1666457, @pirama wrote:

Ping...

@tejohnson Did my reply about shared metadata clarify your question?

It makes more sense to me (note the summary description of the patch needs to be adjusted similarly, it still refers to the summary). I have a couple more questions below.

llvm/lib/Linker/IRMover.cpp

765

In the test SGV and DGV are @d?

llvm/test/LTO/X86/Inputs/type-mapping-bug3.ll

24

"refers to by" doesn't make sense. Do you mean "refers to !5 in..."?

llvm/test/LTO/X86/type-mapping-bug3.ll

4

It's a little odd to say that the copy of d from this module is prevailing, since it is just a declaration. Can this be "%t1.o,d," and would that still trigger the error without your fix?

21

"when its type is mapped to itself incorrectly"?

[pirama]

In D66814#1666487, @tejohnson wrote:

In D66814#1666457, @pirama wrote:

(note the summary description of the patch needs to be adjusted similarly, it still refers to the summary).

Aah, phabricator is not showing my updated commit message. I'll manually update the summary.

llvm/lib/Linker/IRMover.cpp

765

Yes.

llvm/test/LTO/X86/type-mapping-bug3.ll

4

Yes, the error is triggered by %t1.o,d,. I've updated the test to %t1.o,d, since that seems the preferred input.

[pirama]

• Do not have a declaration as a prevailing symbol in llvm-lto2's arguments
• Fix comments in test

[tejohnson]

lgtm

llvm/lib/Linker/IRMover.cpp

for (GlobalValue &SGV : SrcM->globals()) {
 
     // Unify the element type of appending arrays.
     ArrayType *DAT = cast<ArrayType>(DGV->getValueType());
     ArrayType *SAT = cast<ArrayType>(SGV.getValueType());
     TypeMap.addTypeMapping(DAT->getElementType(), SAT->getElementType());
   }
 
   for (GlobalValue &SGV : *SrcM)
-    if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
+    if (GlobalValue *DGV = getLinkedToGlobal(&SGV)) {
+      if (DGV->getType() == SGV.getType()) {
+        // If the types of DGV and SGV are the same, it can only mean that DGV
+        // is from the source module and got added to DstM from a module
+        // summary.  We shouldn't map this type to itself in case the type's
+        // components get remapped to a new type from DstM (for instance, during
+        // the loop over SrcM->getIdentifiedStructTypes() below).
+        continue;
+      }
+
       TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
+    }
 
   for (GlobalValue &SGV : SrcM->aliases())
     if (GlobalValue *DGV = getLinkedToGlobal(&SGV))
       TypeMap.addTypeMapping(DGV->getType(), SGV.getType());
 
   // Incorporate types by name, scanning all the types in the source module.
   // At this point, the destination module may have a type "%foo = { i32 }" for
   // example.  When the source module got loaded into the same LLVMContext, if

llvm/test/LTO/X86/Inputs/type-mapping-bug3.ll

+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; T2 is the non-opaque struct required to trigger the uniqued T2.0 and T3.0 to
+; respectively T2 and T3 in the destination module.
+%"T2" = type { %"T3"* }
+%"T3" = type opaque
+
+; Use/refer to T2 so it gets added as an IdentifiedStructType.  The debug reference to !6 is required to 
+define void @a(%"T2") !dbg !6 {
+  ret void
+}
+
+!llvm.module.flags = !{!0, !1}
+!llvm.dbg.cu = !{!2}
+!0 = !{i32 1, !"ThinLTO", i32 0}
+!1 = !{i32 2, !"Debug Info Version", i32 3}
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !3, retainedTypes: !4)
+!3 = !DIFile(filename: "f2", directory: "")
+!4 = !{!5}
+
+; This DICompositeType refers to by !5 in type-mapping-bug3.ll
+!5 = !DICompositeType(tag: DW_TAG_structure_type, flags: DIFlagFwdDecl, identifier: "SHARED")
+
+!6 = distinct !DISubprogram(unit: !2)

llvm/test/LTO/X86/type-mapping-bug3.ll

+; RUN: opt -module-summary -o %t0.o %S/Inputs/type-mapping-bug3.ll
+; RUN: opt -module-summary -o %t1.o %s
+; RUN: llvm-lto2 run -o %t2 %t0.o %t1.o -r %t0.o,a,px -r %t1.o,b,px -r %t1.o,c,px -r %t1.o,d,px
+;
+; Test for the issue described in https://bugs.llvm.org/show_bug.cgi?id=40312
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; T2 is the non-opaque struct required to trigger the uniqued T2.0 and T3.0 to
+; respectively T2 and T3 in the destination module.
+%"T2" = type { %"T3"* }
+%"T3" = type opaque
+
+; Use/refer to T2 so it gets added as an IdentifiedStructType.
+define void @c(%"T2") {
+    unreachable
+}
+
+; The global declaration that causes the assertion.
+declare void @d(%"T3"*)
+
+define void @b() {
+entry:
+  %f.addr = alloca %"T3"*load %"T3"*, %"T3"** %f.addr
+
+  ; The call with the getCalledValue() vs getCalledFunction() mismatch.
+  call void @d(%"T3"* %0)
+  unreachable
+}
+
+!llvm.dbg.cu = !{!2}
+!llvm.module.flags = !{!0, !1}
+!0 = !{i32 1, !"ThinLTO", i32 0}
+!1 = !{i32 2, !"Debug Info Version", i32 3}
+!2 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !3, retainedTypes: !4)
+!3 = !DIFile(filename: "f1", directory: "")
+!4 = !{!5}
+
+; This DICompositeType is referenced by !5 in Inputs/type-mapping-bug3.ll
+; causing the function type in !11 to be added to its module.
+!5 = !DICompositeType(tag: DW_TAG_structure_type, templateParams: !6, identifier: "SHARED")
+!6 = !{!7}
+
+; The reference to d and T3 that gets loaded into %t0.0
+!7 = !DITemplateValueParameter(value: void (%"T3"*)* @d)